6 Conclusions

We have investigated a new methodology for studying recognition problems | rst we construct a realistic generative model, then we empirically study the resulting distributions to estimate tradeoos between computational requirements and the probability of error. Our experiments show that such simulations can provide a rich source of statistics to guide designers of classiiers, providing worst-case bounds on the computational resources required to achieve speciic error and reject rates. There are many open problems: estimators that can yield tighter bounds should be sought; experiments on more classes, more fonts, and larger images are needed; the defect model must be more carefully validated; and a sampling procedure providing a variance structure that facilitates statistical extrapolation would be helpful. In this methodology, domain knowledge is expressed in the image defect model, and thus is eeec-tively separated from the methodology for designing classiiers. This ooers an escape from the severe practical diiculties that result from the lack of real training data sets of adequate size. We have shown in this paper a successful application of this methodology to a challenging character recognition problem.